ONESTRUCTION builds Ishigaki-IDS BIM foundation model with AWS GenAIIC

ONESTRUCTION, a construction technology startup, built Ishigaki-IDS, a foundation model (FM) specialized for the construction industry's BIM (Building Information Modeling) workflows, with technical advisory from the AWS Generative AI Innovation Center as part of GENIAC Phase 3. The model is based on Qwen3 and addresses three challenges: data scarcity for the new IDS standard, injecting an IFC vocabulary of several thousand terms, and IDS-specific grammar. To overcome these, ONESTRUCTION used synthetic data generation with domain experts, and a three-stage training pipeline: continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning with verifiable rewards (RLVR). For RLVR, they used buildingSMART's IDS-Audit-Tool as a reward function to ensure XML well-formedness, IDS structural validity, and semantic consistency. They trained on Amazon EC2 P5en instances with AWS ParallelCluster and stored data on Amazon FSx for Lustre. In evaluation using their own IDS-Bench benchmark, Ishigaki-IDS scored close to 100 percent on XML and IDS structural compliance and above 80 percent on IDS content consistency, while general frontier models scored under 25 percent on structural compliance and near 0 percent on content consistency. The model supports context-length scaling with YaRN, generating correctly with inputs and outputs up to roughly 120k tokens. A joint proof-of-concept with buildingSMART received positive feedback from both IDS specialists and non-specialists. Lessons learned include the importance of synthetic data quality over quantity, verifiable rewards accelerating iteration, and stable infrastructure allowing free experimentation.